ASA Connect

Dear Colleagues, 

The ASA Statistical Learning and Data Science Section is pleased to announce its next webinar on April 2, featured by Prof. Irina Gaynanova from University of Michigan.  Prof. Gaynanova will discuss about distributional learning for data from wearable devices, with application to continuous glucose monitoring.  Hope to see you there!  

Title:                        Distributional learning for wearable data: methods and applications with continuous glucose monitoring data

Speakers:               Prof. Irina Gaynanova,  Department of Biostatistics, University of Michigan

Date and Time:      April 2, 2025, 3:00 to 4:30 pm Eastern Time

Registration Link:  ASA SLDS Webinar Registration Link [eventbrite.com] 

Abstract:                Continuous glucose monitors (CGMs) provide frequent glucose measurements over time, offering valuable insights for diabetes management. Traditionally, these data are represented by percentages of time spent within fixed thresholds or simple statistics like mean glucose, which can lead to information loss. Distributional learning provides a more comprehensive framework for data representation by utilizing the entire distribution of glucose measurements, thus enabling improved data interpretation and analysis. This talk presents two applications of distributional learning for CGM data: optimizing threshold-based summaries and selecting significant covariates in distributional regression. First, we examine the limitations of fixed glucose thresholds (e.g., 70–180 mg/dL) and introduce a data-driven, loss-based approach that optimizes thresholds by preserving key distributional properties. Using the Wasserstein distance as the base measure, we reformulate the loss minimization as optimal piecewise linearization of quantile functions. Applying this method to CGM data from individuals with type 1 diabetes and those with normal glycemic control, we show that data-driven thresholds vary by population and improve discriminative power over fixed thresholds. Next, we present a new optimization algorithm for sparse distributional regression, enabling efficient variable selection and inference via stability selection. Applying this method to CGM data from individuals with type 2 diabetes, we identify significant associations between glucose variability and medication use, independent of mean glucose levels-demonstrating the advantages of a distributional approach in capturing dependencies that extend beyond the first moment. While our primary focus is CGM data, these methods extend to other types of high-frequency wearable data, such as data from actigraphy monitors.

Presenter:              Irina Gaynanova is an Associate Professor in the Department of Biostatistics at the University of Michigan. She received her PhD in Statistics from Cornell University in 2015. Dr. Gaynanova's teaching emphasizes reproducible research practices, statistical computing, and communication skills to prepare students for STEM careers. Her research focuses on developing statistical methods for analyzing high-dimensional biomedical data. Her methodological interests include data integration, machine learning, and high-dimensional statistics, driven by challenges in multi-omics analyses and wearable device data, particularly continuous glucose monitors. Her research has been funded by the National Science Foundation and the National Institutes of Health, and she served on the editorial boards of the Annals of Applied Statistics, Biometrika, Data Science in Science, the Journal of the American Statistical Association, and the Journal of Computational and Graphical Statistics. Her contributions to research, teaching, mentorship, and service have been recognized with a David P. Byar Young Investigator Award, an NSF CAREER Award, elected membership to the International Statistical Institute, the IMS Thelma and Marvin Zelen Emerging Women Leaders in Data Science Award and the COPSS Emerging Leader Award.